BENCHMARKS · GPU UTILIZATION
GPU Economics
Pick a model, an accelerator, a datacenter region, and a price — and find the utilization where the node turns a profit.
Model
The reference scenario from the post
Accelerator
NVIDIA · 288GB HBM3e
Datacenter region
The dollar premium — power, land, labor, and the queue for capacity
Serving stack
GPUs per node
GPU rent · $/chip-hr
Pricing
Break-even utilization32%at $1.00 per 1M output tok
Node throughput35,000output tok/s · sglang
Revenue vs rent$63 / $40per hour at 50% util
Profit / month$16,790after GPU rent only
Revenue vs utilization — DeepSeek V4 Pro on 8×B300 · United States
Same model, every accelerator — DeepSeek V4 Pro · sglang · off-peak API price
| Accelerator | Rent $/hr | tok/s | Rev @100% | Break-even | Profit/mo @50% |
|---|---|---|---|---|---|
| H200 NVIDIA | $44 | 3,150 | $11 | never | -$28,039 |
| B200 NVIDIA | $40 | 38,300 | $138 | 29% | $21,126 |
| B300 NVIDIA | $40 | 35,000 | $126 | 32% | $16,790 |
| MI300X AMD | $26 | 9,500 | $34 | 76% | -$6,497 |
| MI355X AMD | $34 | 16,000 | $58 | 59% | -$3,796 |
| TPU v7 Google | $24 | 6,500 | $23 | never | -$8,979 |
GPU rent only — before bandwidth, CPU head nodes, power overhead, staff, and every other line item finance will ask about. Serving throughputs and regional rent premiums are illustrative figures calibrated to published ballparks, not a live feed.